Google launches Gemini 3 and the perceived capability gap with OpenAI and Anthropic effectively closes
Two model sizes, benchmark scores at parity with GPT-5 and Claude 4, and a serving infrastructure that runs entirely on Google's TPU stack. The technical question is largely settled. The commercial question is not.

On 12 November 2025, Google released Gemini 3 in two sizes, Pro and Flash, with a Pro Deep Think variant available to top-tier subscribers. The benchmark headline numbers, reported in the launch post and corroborated by independent reviews from MIT Technology Review and The Verge, sat at parity with GPT-5 and Claude 4 across the standard public evaluations. On some reasoning-heavy benchmarks, Gemini 3 Pro Deep Think scored above both.
The launch was Google's first frontier-AI release since the spring 2024 rollout of Gemini 1.5 to be uniformly read as competitive on capability. The intervening eighteen months of Gemini 2 family releases had been received as solid but trailing. The Gemini 3 family, on the technical evidence, closed that perceived gap.
Benchmarks at launch
The Pro Deep Think variant was particularly striking on ARC-AGI 2 and the Humanity's Last Exam benchmark released earlier in 2025. The more interesting structural detail was that Gemini 3 served entirely on Google's TPU infrastructure, including the v6 generation that had reached general availability in mid-2025. The implied independence from Nvidia compute became, in the days following launch, a separate strand of analyst commentary.
The commercial question
The capability question is largely settled in Gemini 3's favour, at least at parity. The commercial question is not. As reported by Bloomberg in coverage of the launch, Google's enterprise-AI revenue going into Q4 2025 was running at an estimated annualised rate behind both OpenAI and Anthropic, despite a much larger installed base of Google Cloud customers. Whether the technical-capability narrative converts into developer-platform share, in the way OpenAI and Anthropic have built theirs, was the live commercial question heading into 2026.
Gemini 3 closed the technical gap. The developer-platform gap is harder.
Internal Alphabet messaging following the launch, captured in coverage by The Information, framed the next twelve months as primarily a developer-relations and enterprise-product effort rather than a model-research effort. The model-research lead, in Google's own framing, was no longer the binding constraint.



Discussion